Fine-tuning vs RAG: the criterion
Default: RAG when knowledge must stay current and citations matter. Fine-tune when style, language or a narrow skill doesn't fit in context. Hybrid when both are true.
Fine-tuning and RAG solve different problems. This is the decision order we follow before spending a peso on training.
RAG first, almost always
Changing knowledge + required citations = RAG.
If your knowledge changes (catalogues, policies, documentation) or your answers must carry citations, RAG is the default: index, re-rank, cite — without touching the model.
RAG fails when the problem isn't knowledge but form: tone, output format, or a narrow task where the generic model's quality falls short of your bar.
Fine-tune for form, not facts
LoRA adapters for tone and format; facts stay in RAG.
Fine-tuning wins when you need the model to sound, format or execute like your organization: editorial style, fixed output structures, stable narrow domains.
We use LoRA adapters — we don't retrain the world. Cost stays bounded, rollback is possible, and knowledge keeps living in RAG where it can be updated.
Hybrid, when both are true
Adapters for form + RAG for knowledge.
When style must be yours and knowledge must stay current, hybrid wins: adapters for form, RAG for data. It's the configuration that survives production best.
Four questions before paying for training
Answer in writing; the route follows.
Before paying for training, answer in writing: is the problem knowledge or form? Do your data change daily or stay stable? Must citations be verifiable? Must knowledge stay in-house? The answers define the route; everything else is consultant margin.
Key takeaways
- RAG by default: changing knowledge and required citations.
- Fine-tune only for form: style, format, narrow tasks.
- Hybrid when both are true.
- Nothing is decided without a measured baseline.
- Local inference when compliance demands it.
The questions we hear often
When is fine-tuning NOT worth it?
When the problem is missing or changing knowledge — RAG handles that cheaper and safer.
What's a LoRA adapter?
A small trained layer over the base model for form and narrow tasks: cheap, reversible, local-inference friendly.
Can RAG and fine-tuning combine?
Yes — that's the common production pattern: adapters for form, RAG for cited knowledge.
How is the outcome measured?
Against a generic baseline with comparable quality, cost and latency metrics.
Does this improve Spanish?
We evaluate business Spanish explicitly; it's a lab axis.
Where to next
Fine-tuning & RAG
When to tune the model and when to give it knowledge. Judgement, not hype.
Learn more→ AI LabSlash AI Lab
Our lab where we evaluate models and publish the method.
Learn more→ ModelosHow we choose a model
Evaluation in Spanish and French. Production criteria.
Learn more→ AIArtificial intelligence
Applied AI for real operations: agents, RAG, vision, voice and automation.
Learn more→Let's put it in production
Tell us your challenge. We reply within 24 hours with an honest first read: if we can help, we'll say how; if not, we'll say who can.
I reply personally. No endless forms, no canned replies.